Inspiration : The idea was inspired by how industries and engineers still rely on manual visual inspections to find defects or damages.
In fields like manufacturing, infrastructure monitoring, and F1 car design, even a small scratch or change can have a huge impact. We wanted to automate this process using AI and computer vision, making it faster, safer, and more reliable.
What it does : The Visual Difference Engine automatically compares two or more images taken at different times and detects what has actually changed between them It:
Highlights real differences such as cracks, scratches, missing parts, or shape/design changes.
Ignores small variations caused by lighting, shadows, or camera angle.
Produces an output image that marks the changed regions clearly.
Can be used for automated inspections, monitoring infrastructure, or tracking design updates.
How we built it : 1. Preprocessing:
We first aligned two images using ORB keypoint matching to correct small position differences.
Change Detection: We used image differencing with OpenCV for a baseline and later trained a Siamese U-Net to detect complex changes.
Post-processing: We removed noise using morphological filters and highlighted real changes with bounding boxes.
Visualization: The output shows changed regions clearly, helping users quickly spot defects.
Challenges we ran into : Lighting differences caused false detections — solved using histogram matching.
Camera angle mismatch — fixed with image alignment techniques.
Small dataset size — created synthetic images with artificial defects to train the model.
Performance on low-end devices — optimized using quantized models and efficient libraries.
Accomplishments that we're proud of : Successfully built a working AI system that accurately detects and highlights changes.
Created a synthetic dataset generator to train the model effectively.
Achieved high precision while ignoring lighting and angle differences.
Designed a simple, interactive dashboard for real-time use.
What we learned : Through this project, we learned how to:
Align and compare two time-series images using OpenCV and feature matching.
Apply deep learning models (like Siamese U-Net) for image difference detection.
Handle challenges like lighting variations, camera angle changes, and false detections.
Build a small end-to-end system — from preprocessing to model deployment.
We also explored some image similarity concepts like
D(x, y) = |I_1(x, y) - I_2(x, y)|
What's next for Visual difference Engine - concise solution : Add AI-based classification to categorize change types (e.g., crack, dent, rust).
Connect to IoT cameras for real-time industrial and infrastructure monitoring.
Deploy on cloud dashboards for centralized monitoring and automated alerts.
Extend for medical or satellite image change detection in future versions.
Log in or sign up for Devpost to join the conversation.